ABCP 2026 Annual Conference: Artificial Intelligence and Cyber Security Track

Yulan He

Abstract:

Two major advances in large language models, namely, better reasoning and reward-based reinforcement learning, can create unexpected safety risks. First, we show that improving a model’s reasoning ability can make it less safe, a problem we call Reasoning-Induced Misalignment. We explain why this happens through mechanistic interpretability analysis and find that fine-tuning with chain-of-thought reasoning traces can mix reasoning abilities with safety-related mechanisms, causing the model to forget previously learned safeguards. Second, we show that reward hacking can extend beyond technical tasks into social and regulatory settings. Using our constructed SocioHack environments, we find that models can learn strategies that follow rules on paper while undermining their intended purpose. Existing safeguards provide only limited protection against this behaviour. Together, these findings show that improvements in reasoning capability and reward-based reinforcement learning can expose gaps in both model internals and human-designed rules. This highlights the need for safer post-training methods guided by mechanistic interpretability.

Bio:

Yulan He is a Professor in Natural Language Processing at King’s College London and a Turing AI Fellow. Her research focuses on improving the robustness of Large Language Model (LLM) reasoning, agentic AI, long-context QA, model interpretability, safety alignment, AI for education, health, and science. She has received several prizes and awards for her research. Her five-year Turing AI Fellowship project, “Event-Centric Framework for Natural Language Understanding”, was awarded Best Research Project (Research Excellence) at the RAi UK AI & Robotics Awards 2026. She also received a Best Research Paper Award at the same awards ceremony. In addition, she is the recipient of a SWSA Ten-Year Award and a CIKM Test-of-Time Award, and was named an inaugural Highly Ranked Scholar by ScholarGPS.

Jindong Gu

Abstract:

AI agents are turning foundation models into software that can perceive their environment, reason over uncertain information, use tools, and autonomously take actions. This changes cybersecurity: application behavior is no longer determined solely by developer-written code, but is partly generated at runtime by an autonomous decision-maker. In this talk, I will explore how agents expand the security boundary beyond code and APIs to include context, decisions, action trajectories, and interactions among agents. I will discuss emerging vulnerabilities such as indirect prompt injection and manipulation of multi-agent decision making, and present our recent work on reliable tool use and agent security. I will conclude with a broader question: when software can make decisions, what does it mean to secure the software?

Bio:

Jindong Gu is a Senior Researcher at the University of Oxford. His research focuses on building Responsible AI, with current interests in the reliability, safety, and security of AI agents and foundation models. He has published 100+ papers with 10k+ citations, regularly served as Area Chairs, and organized workshops at top AI conferences. His work has received multiple honors, including Best Paper Awards at IJCAI 2025 and CVPR 2025 workshops, OpenAI Red-Team Challenge Honorable Mention, AI Rising Star Award at KAUST.

Abstract:

Diffusion models can generate biased or unsafe images, yet why they do so remains unclear from the perspective of their internal representations. This talk presents a self-supervised approach for discovering interpretable latent directions for arbitrary concepts in diffusion models, and shows how these directions can be used as a simple, training-free intervention to improve fairness, safety, and responsible generation. The talk closes with broader implications for interpretability as a route to controllable generative models.

Bio:

Hang Li is a Research Scientist at Meta AI in London, where he works on generative AI with a focus on diffusion and video foundation models, multimodal conditioning, and controllable generation. He received his Ph.D. in Computer Science from LMU Munich under the supervision of Prof. Volker Tresp, affiliated with the Munich Center for Machine Learning and Siemens AG, and holds an M.Sc. from the Technical University of Munich. His work has appeared at CVPR, ICCV and WACV, spanning interpretability and responsible generation in diffusion models, multimodal reasoning, and controllable image and video synthesis.

Yukun Lai

Abstract:

In recent years, deep generative models have significantly reduced the effort needed for visual content creation, with only minimal user input such as text prompt. However, there are often ambiguities and automatically generated visual content may not follow users’ intention. While it is possible to re-generate the content, e.g. with modified text prompts,  it cannot ensure unedited regions are fully retained, and it cannot achieve detailed control. In this talk, I will present some of our recent works that aim to provide intuitive, detailed control for visual content generation and editing, including images, videos and 3D shapes, where intuitive user control such as scribbles and sketches are used to ensure fine-grained control, while maintaining quality and efficiency.

Bio:

Yukun Lai is a Professor in the School of Computer Science and Informatics, Cardiff University since 2020. He received his bachelor’s and PhD degrees from Tsinghua University in 2003 and 2008, respectively. He is the Cardiff lead of the EPSRC AI Hub in Generative Models and co-leads Multimodal Models working group. His research is in the areas of Computer Graphics, Geometric Processing, and Computer Vision. He is an associate editor of IEEE Transactions on Visualization and Computer Graphics, The Visual Computer and The European Journal of Artificial Intelligence. He has published more than 100 papers in leading journals and conferences, with over 17K citations. He is a fellow of the Learned Society of Wales.

Hongpeng Zhou

Abstract:

Foundation models are transforming biomedical research by learning generalisable representations from large-scale data, yet their application to spatial transcriptomics remains underexplored. In this talk, I will introduce NexuST, a next-generation hierarchical transformer-based foundation model that jointly captures gene-expression programmes and the spatial organisation of cells. NexuST enables gene-level and cell-level representations to exchange information throughout training, allowing intracellular gene-expression states and local microenvironments to iteratively refine one another. The model was pretrained on HumanST-46M, comprising approximately 46 million cells from 72 datasets, 123 tissue slides, 11 organs and three imaging-based platforms. Across held-out datasets containing approximately 2.6 million cells, NexuST achieved state-of-the-art or competitive performance in cell-type annotation, tissue-region prediction, gene-expression recovery and neighbourhood-composition prediction, demonstrating the potential of hierarchical foundation models to advance spatially resolved biological discovery.

Bio:

Hongpeng Zhou is a Dame Kathleen Ollerenshaw Fellow (Assistant Professor/Lecturer) at the Computer Science Department, University of Manchester, UK. He obtained his PhD from Delft University of Technology in 2022. His research focuses on machine learning and AI for Biomedicine. He earned his Bachelor’s and Master’s degrees from Harbin Institute of Technology, China.

Jinwei Hu

Yi Dong

Abstract:

As large language models (LLMs) transition to autonomous agents synthesizing real-time information, their reasoning capabilities introduce an unexpected attack surface. This paper introduces a novel threat where colluding agents steer victim beliefs using only truthful evidence fragments distributed through public channels, without relying on covert communications, backdoors, or falsified documents. By exploiting LLMs’ overthinking tendency, we formalize the first cognitive collusion attack and propose Generative Montage: a Writer-Editor-Director framework that constructs deceptive narratives through adversarial debate and coordinated posting of evidence fragments, causing victims to internalize and propagate fabricated conclusions. To study this risk, we develop CoPHEME, a dataset derived from real-world rumor events, and simulate attacks across diverse LLM families. Our results show pervasive vulnerability across 14 LLM families: attack success rates reach 74.4% for proprietary models and 70.6% for open-weights models. Counterintuitively, stronger reasoning capabilities increase susceptibility, with reasoning-specialized models showing higher attack success than base models or prompts. Furthermore, these false beliefs then cascade to downstream judges, achieving over 60% deception rates, highlighting a socio-technical vulnerability in how LLM-based agents interact with dynamic information environments. Our implementation and data are available at: https://github.com/CharlesJW222/Lying_with_Truth/tree/main.

Bio:

Jinwei Hu is a PhD researcher in School of Computer Science and Informatics at the University of Liverpool and a member of the Trustworthy Autonomous Cyber-Physical Systems (TACPS) Laboratory. His research focuses on Trustworthy and Responsible Agentic AI, with particular interests in the safety, reliability, and robustness of LLM-driven agents and multi-agent systems operating in dynamic and adversarial environments. His research interests include AI safety and security, multi-agent risks and belief manipulation, fine-grained control of LLMs, testing and verification of AI-enabled systems, and AI4Science. His work has been published in leading venues including ACL, NeurIPS, AAAI, ICML, CVPR, TACL, and IEEE Transactions, with his recent work on multi-agent collusion and belief manipulation receiving an Outstanding Paper Award at ACL 2026.

Yi Dong is Assistant Professor at the Department of Artificial Intelligence, University of Liverpool. He co-leads the TACPS (Trustworthy Autonomous Cyber Physical Systems) Lab, focusing on responsible AI systems. He was a former New Frontiers Fellow at the University of Southampton. He has published 40+ papers at top venues like AAAI and NeurIPS, and has secured over £200K in UKRI and Alan Turing Institute funding.

Zhipeng Wang

Abstract:

Autonomous AI agents are beginning to act as economic participants: discovering services, deciding whether unfamiliar counterparts can be trusted, and making payments without human approval. An emerging agentic Internet stack is taking shape, including A2A and MCP for communication, ERC-8004 for identity and reputation, and x402 for web-native payments. But can agents safely rely on this infrastructure when real value is at stake?

This talk presents two recent studies of its trust and payment layers. We first examine x402 and show how the gap between web authorization and blockchain settlement leads to practical security failures, including replay attacks, settlement inconsistencies, and manipulation of service discovery. We then study ERC 8004 across Ethereum, BSC, and Base, finding that many registered identities are inactive and that reputation signals are manipulable and susceptible to Sybil behavior. Together, these results show that agents can already pay, but the infrastructure for deciding whom to trust and when payment is safely complete is not yet trustworthy. This talk also concludes with practical recommendations for building more secure agent economies.

Bio:

Zhipeng Wang is a Lecturer (Assistant Professor) in Cyber Security in the Department of Computer Science at the University of Manchester. He was previously a Postdoctoral Research Associate at Imperial College London, where he also completed his PhD in Computing. His research focuses on security, privacy, and trust in open and distributed systems, spanning applied cryptography, blockchain, and emerging agentic AI systems. He has published at top security and privacy venues such as IEEE S&P, USENIX Security, NDSS, WWW, and TIFS. His work is supported by prestigious industrial grants, including the Ethereum Foundation Academic Grant.

Charlie Fu

Abstract:

Artificial intelligence (AI) is increasingly being applied across the built environment; however, its effective use in architectural design remains limited. This presentation introduces a suite of AI applications developed to automate the design of school buildings. It begins by presenting an innovative method for evaluating existing AI applications in building design and planning, together with the results of an assessment of 50 tools available online. It then explains the design and development of an AI-based tool for educational building design, focusing on two key considerations: the functions and dimensions of spaces, and the spatial relationships between them. The tool uses machine learning (ML) and large language model (LLM) technologies—specifically Claude and Claude Code—to develop its understanding of key design principles. Then, the presentation demonstrates how the tool uses an evolutionary algorithm (EA) to automatically generate floor plans that respond to site boundaries. It also shows how Rhino.Inside.Revit is used to 3D Building Information Modelling (BIM) design models incorporating essential horizontal and vertical circulation spaces. Although the system is developed using existing AI and BIM technologies, it demonstrates an effective integrated approach to automating school-building design, with the potential for adaptation to other building types.

Bio:

Charlie Fu is Professor of the Built Environment at the University of West London (UWL). He joined UWL in 2009 and previously served as Subject Leader for the Built Environment. He has developed several undergraduate and postgraduate programmes and established a research group focusing on building sustainability and digitalisation. Professor Fu is also a Fellow of the Chartered Institute of Architectural Technologists (FCIAT), a Chartered Member of the Royal Town Planning Institute (MRTPI), and a member of the EPSRC Peer Review College. Before joining UWL, he worked at Salford University and Cambridge University.

Bo Wei

Abstract:

This talk is about how edge AI can deliver fast, privacy-aware intelligence directly on devices while addressing the security challenges of distributed, resource-constrained deployments. It highlighted practical approaches to building robust, efficient, and trustworthy edge systems.

Bio:

Bo Wei is a Senior Lecturer in the School of Computing at Newcastle University. He is also a member and Co-PI of an EPSRC-funded national Edge AI Hub. Previously, he was a Lecturer at Lancaster University and a Postdoctoral Research Assistant at Oxford University. He obtained his PhD in Computer Science and Engineering from the University of New South Wales, Australia, in 2015. His research interests include Edge AI, the Internet of Things, and cybersecurity.

Xinqi Fan

Abstract:

Ulcerative colitis is a chronic inflammatory bowel disease that causes recurrent inflammation in the colon and requires careful endoscopic assessment for diagnosis. Artificial intelligence is rapidly reshaping how we analyse endoscopic data in ulcerative colitis, moving beyond score prediction tasks toward richer and more clinically meaningful understanding. In this talk, I will present our recent work across this progression. I will first discuss discriminative or encoder-based models based on convolutional neural networks and transformer architectures for automated severity scoring, focusing on clinically important indices. I will then introduce our more recent exploration of multimodal large language models using a mixture of experts for ulcerative colitis captioning, where visual findings are translated into natural language descriptions. Together, these studies illustrate a broader shift from scoring disease severity to enabling multimodal interpretation, and highlight the promise of building more expressive and clinically useful AI systems for endoscopic analysis.

Bio:

Xinqi Fan is a Lecturer in Artificial Intelligence at Manchester Metropolitan University. He received his BEng from Southwest University, his MEng from the University of Western Australia, and his PhD from City University of Hong Kong. He also conducted research at King Abdullah University of Science and Technology and the Chinese University of Hong Kong. His research interests include deep learning, computer vision and multimodal learning, with applications in affective computing and medical image analysis. He has published research in leading conferences and journals, including CVPR, ICCV, MM, IEEE TAFFC and IEEE TIP. He has organised challenges at FG 2026 and MM 2025, as well as a workshop at ICME 2025. His team won first place in the ISBI 2026 Multimodal Ulcerative Colitis Grading Challenge.

Gaoshan Bi

Abstract:

Accurate terrestrial carbon flux prediction is essential for understanding how ecosystems function as carbon sources and sinks. However, existing deep-learning models often struggle to generalise across heterogeneous biomes. A key challenge is ecological memory: current carbon exchange can continue to be influenced by environmental conditions from previous weeks or months, while the relevant timescales vary substantially across ecosystems.

This talk presents MemPerceiver, a multimodal framework that combines short-context meteorological and satellite observations with adaptive, long-horizon ecological memory. It summarises antecedent conditions across six time horizons ranging from 7 to 180 days and uses a sample-conditioned gating mechanism to learn which historical horizons are informative for each prediction. Evaluated on the CarbonSense benchmark, comprising 385 eddy-covariance sites worldwide and spanning 15 ecosystem classes, MemPerceiver improves the prediction of Net Ecosystem Exchange (NEE) and Gross Primary Production (GPP), with particularly clear gains in data-scarce and previously unseen ecosystems. Beyond predictive performance, the learned gate activations form interpretable temporal fingerprints that reveal both shared and biome-specific patterns of ecological memory. Overall, this work demonstrates how adaptive multi-scale memory can improve cross-ecosystem generalisation while generating model-derived hypotheses about how different ecosystems respond to antecedent environmental conditions.

Bio:

Gaoshan Bi is a third-year PhD researcher and Research Assistant at the University of Sheffield, supervised by Professor Po Yang. His research lies at the intersection of multimodal deep learning, earth observation and environmental science. He develops AI methods that integrate meteorological, satellite and sensor data for applications including carbon flux prediction and climate-smart agriculture.

Kewen Sun

Abstract:

As a core component of Positioning, Navigation, and Timing (PNT) systems, Global Navigation Satellite Systems (GNSS) deliver continuous, high-precision positioning, velocity, and timing services to users in near-Earth space and on the ground. Their role has become indispensable across both military and civilian domains, underpinning critical infrastructure, transportation, scientific research, social production, and everyday activities. Nevertheless, the growing dependency on satellite navigation has also exposed notable vulnerabilities and emerging security threats. Due to the extremely low received power of GNSS signals and the unencrypted nature of civilian frequency bands, GNSS receivers are particularly susceptible to both intentional and unintentional electromagnetic interference, which can severely impair GNSS signal acquisition, tracking, and overall navigation performance. This talk will examine recent advances in the detection, characterization, and mitigation of GNSS interference through transformed-domain signal processing techniques. Emphasis will be placed on advanced time-frequency analysis methods, including the Fractional Fourier Transform, Wavelet Transform, Radon-Wigner Transform, and Hough Transform-based anti-interference strategies. These techniques provide robust analytical frameworks for revealing distinctive interference features and improving the detection and mitigation of diverse interference types in complex, contested electromagnetic environments.

Bio:

Kewen Sun is an Associate Professor of Space Technologies at the Bradford-Renduchintala Centre for Space AI, University of Bradford, U.K. His research interests include satellite navigation and communications, wireless communications, space AI, and advanced signal processing, with a particular focus on GNSS interference detection and mitigation. He received his Ph.D. in Electronic and Communication Engineering from Politecnico di Torino, Italy, in 2010. Before joining the University of Bradford, he was a Professor at Hefei University of Technology, China, and a Visiting Professor at Imperial College London, U.K., from November 2019 to November 2020. He currently leads ESA-funded research on Joint Communications and Positioning Anti-Jamming Beamforming. He has published in leading journals and conferences, including IEEE TAES, IEEE TVT, IEEE TIM, ION GNSS+, ION ITM, and IEEE/ION PLANS. He is also an inventor on several patents in GNSS weak-signal processing and anti-interference technologies and serves on the Youth Editorial Board of Satellite Navigation.

Po Yang

Abstract:

Agriculture is facing increasing pressures from climate change, resource constraints, environmental challenges, and the need to maintain productivity and profitability. Artificial intelligence (AI) offers significant opportunities to support more efficient and sustainable agricultural systems but translating AI from laboratory models into real-world farming practice remains a major challenge. Agricultural environments are highly dynamic, heterogeneous, and uncertain, therefore AI systems have to be respond to changing conditions, provide understandable recommendations, and operate reliably in practical settings. In this talk, Prof. Po Yang will present a vision for responsive, explainable, and deployable AI for sustainable and resilient agriculture, drawing on a range of UKRI-funded research projects led by his team at the University of Sheffield. The talk will explore how multimodal data, machine learning, computer vision, IoT, and emerging foundation models can be integrated with agronomic knowledge to support data-driven agricultural decision-making. Examples will include AI-driven climate-smart fertiliser management, precision agriculture, pest and disease detection, and plant-parasitic nematode diagnostics.

Bio:

Po Yang is Professor of Pervasive Intelligence in the School of Computer Science at the University of Sheffield, where he leads the Pervasive Computing Group. He is an Independent Scientific Advisor to the Alan Turing Institute, a member of BridgeAI’s Agrifood Expert Group. His research focuses on pervasive computing, mobile sensing, and data intelligence, with applications in proactive health and smart agriculture. Over the past 5 years, he has led more than 20 research projects funded by EPSRC, Innovate UK, and BBSRC, with a combined value exceeding £5.2 million, and has co-managed six EU FP7 projects with funding exceeding €5.6 million. Professor Yang has published more than 200 papers as first or corresponding author in leading AI conferences and IEEE journals, including ESI Highly Cited Papers. His publications have received more than 9,800 citations, with an h-index of 47. He has received five Best Paper Awards, including awards at IEEE INDIN-2019 and the IEEE Industrial Informatics Technical Committee–TII in 2022, and has been recognised among Stanford’s top 2% of scientists globally for three consecutive years since 2022. He serves on the editorial boards of leading journals, including Editor-in-Chief of Array, the IEEE Journal of Translational Engineering in Health and Medicine and the Journal of Biomedical Informatics.

Abstract:

World models aim to learn compact representations of how an environment evolves, enabling prediction, reasoning, and planning without reconstructing every detail of the observed world. In this talk, I will discuss probabilistic world models through the lens of Joint-Embedding Predictive Architectures (JEPAs), with a particular focus on Variational JEPA (VJEPA) and its extensions. While conventional JEPAs typically predict a single future representation, probabilistic JEPAs instead model a distribution over possible future latent states, allowing uncertainty, stochasticity, and multimodal futures to be represented explicitly. I will briefly introduce the probabilistic formulation of VJEPA and discuss several variants that connect JEPA-style prediction with latent-variable models, mixture distributions, and Markovian dynamics. Together, these perspectives suggest a broader view of JEPA as a framework not only for representation learning, but also for uncertainty-aware world modelling.

Bio:

Yongchao Huang is a Lecturer in Computing Science whose research focuses on probabilistic machine learning. His recent work explores world models and Joint-Embedding Predictive Architectures (JEPAs), with particular emphasis on probabilistic latent representations and dynamics. He received PhD from Oxford, and worked in industry, Oxford and Cambridge. He has broad interests in machine learning and applications, and welcome collaborations across domains.

Qiang Fu

Abstract:

The talk will cover four main topics: 1) recent milestones in AI and longevity biotechnology; 2) four foundational pillars of longevity medicine; 3) comparative analysis of the evolution of AI and longevity medicine; and 4) latest research progress integrating AI and longevity science.

Bio:

Qiang Fu is a Visiting Scholar at the Cambridge Stem Cell Institute, University of Cambridge, Professor at the Institute of Ageing Medicine, Shandong Medical and Pharmaceutical University, and Professor at the Innovation Center for Anti-Ageing, Shanghai Jiao Tong University. He also serves as Founder & Executive Dean of the International Longevity Medicine Research Institute and International Longevity Medicine Alliance. His professional affiliations include Chairman of the Longevity Medicine and Anti-Ageing Research Committee under the Shandong Society of Translational Medicine, Director and Member of Expert Committee of the China Anti-Aging Promotion Association, and President of the Yantai Biotechnology Society. He has presided over and completed more than 10 research projects including General Programmes of the National Natural Science Foundation of China and key provincial research programmes, and has published over 30 SCI-indexed papers in prestigious journals such as Nature Metabolism, Aging, Aging Cell, Journal of Experimental Medicine and Oncogene.